Architect-Ant: Editable Automatic Furnishing of Architectural Floor Plans

Architect-Ant introduces a framework for furnishing residential floor plans by treating furniture layout synthesis as structured sequence generation over editable geometric objects, combining pseudo-labeled data with deterministic rule scoring and direct preference optimization to produce geometrically valid and functionally plausible layouts. This method addresses the limitation of existing automatic furniture arrangement by focusing on generating object-level representations in a compact Domain-Specific Language (DSL) that preserves editability, allowing for workflows in real estate visualization and interior design.

Data Preparation and Representation

The research utilizes AntPlan-270, a curated dataset of 270 architectural floor plans with per-room furniture bounding box annotations across ten residential room categories. This dataset is constructed by extracting structural primitives (walls, doors, windows) using an RT-DETR-X detector trained on CubiCasa5K and bootstrapping furniture bounding boxes from a hand-labeled subset. The resulting data is decomposed into room-level samples, where each sample carries the room geometry in metric coordinates and a furniture pseudo-label list with per-instance bounding boxes. The output layout is represented using a compact line-oriented grammar: FURNITURE OBJ class= x= y= w= h=, which allows for direct computation of geometric and combinatorial constraints like overlap, containment, clearance, door obstruction, reachability, wall affinity.

Model Training Pipeline

The training involves a multi-stage process to adapt a pretrained vision-language model (Qwen3.5-9B) to the target representation. The stages include:

1. Prompting for an initial prior over furniture categories and coarse spatial relations.

2. A lightweight fine-tuning stage on pseudo-labeled layouts to adapt the model to the target format and approximate room statistics using pseudo-labeled layouts.

3. A rule-based evaluator that scores sampled layouts using geometric and semantic criteria, such as containment within the room, object overlap, door access, traversable paths, wall affinity, which supplies explicit design preferences.

4. Preference optimization over candidate object placements to further refine layout quality by training the model to assign higher probability to better-scoring layouts.

Inference and Selection

At inference time, Architect-Ant operates through a two-step process: first, the model samples K DSL candidates per prompt using a temperature of 0.9 and top-p of 0.95; second, the deterministic rule scorer ranks these candidates and selects the highest-scoring one as the system output. This ensures that layouts are selected based on explicit geometric and semantic criteria, which are decomposed by rule family to provide a traceable preference signal for training.

Evaluation Metrics

The performance of Architect-Ant is evaluated using two complementary views. The headline view is the per-room mean ± standard deviation of scores across the K candidates per prompt, reflecting typical generation quality and consistency. The secondary view is best-of-K, the average score of the best candidate per prompt selected by the scorer, which reflects the inference-time protocol. Furthermore, an independent visual judge (Gemini 3 Flash Preview) is employed to provide a separate evaluation signal on functional layout quality, focusing on criteria such as wall intersections, door and passageway clearance, functional grouping and wall hugging.

Key Findings

Experiments on out-of-distribution datasets like CubiCasa5K show that Architect-Ant produces geometrically valid and functionally plausible layouts. The synthetic-pair Direct Preference Optimization (DPO) recipe is used as the main recipe, as it localizes each preference to a single perturbed bounding box, whereas broader model-pair DPO can lead to reward hacking by maximizing the rule score without improving visual quality. The VLM-as-judge study indicates that while hard rule scores are useful, softer spatial preferences—such as placing chairs more symmetrically around tables or tightening kitchen groupings—are often captured by the judge, suggesting that DPO improves visual-functional quality in specific categories like the bathroom and kitchen. The system successfully maintains a separation between the editable structured DSL and the downstream blueprint-style visualization, where a domain-specific diffusion model renders the symbolic layout.

The gist: Architect-Ant generates geometrically valid and functionally plausible layouts by training a vision-language model on pseudo-labeled data, using deterministic rule scoring to derive preference signals for direct preference optimization, and selecting the highest-scoring layout at inference time.

How it works

1. The system formulates furnished room layout synthesis as structured sequence generation over editable geometric objects rather than image generation.

2. It adapts a pretrained generator to this representation using pseudo-labeled layouts, providing a task-specific starting point for later preference optimization.

Here are the specific improvements that can be made to existing AI systems, based on the principles and methodology presented in this paper:

The core improvement is shifting from generating raw pixel outputs (images) to generating structured, editable data (a Domain-Specific Language or DSL) that is inherently verifiable and controllable.

1. textbfFrom Image Generation to Structured Layout Synthesis (Architect-Ant Framework):

2. A new AI system can take a simple floor plan geometry and a list of required furniture items, and instead of outputting a final rendered image, it outputs an editable text structure (DSL) describing the exact position, size, and type of every object. This DSL is the source of truth for all subsequent architectural workflows.

3. textbfFrom Black-Box Generation to Constraint-Driven Learning (Rule-Based Scoring):

4. The AI's learning process is augmented by a deterministic rule-based evaluator that assigns explicit penalties based on geometric and semantic constraints (e.g., wall clearance, door swing, object overlap). This moves the system beyond purely statistical accuracy to ensure functional validity.

5. textbfFrom Post-Hoc Repair to Training Supervision (Preference Optimization via DPO):

6. The AI is trained not just on correct outputs, but on a preference signal derived from the rule scorer. By using Direct Preference Optimization (DPO), the model learns to assign higher probability to layouts that satisfy these explicit geometric and functional constraints during training, rather than relying solely on external checks applied after generation.

7. textbfFrom Single-Objective Learning to Multi-Signal Supervision (Procedural Reasoning Traces):

8. The training data is enriched with procedural reasoning traces that show the step-by-step decision-making process for placing each object, including in some cases, on-the-fly correction of a placement error. This provides the model with a rich supervisory signal about architectural constraints (like wall affinity and circulation) that simple image labels cannot convey.

9. textbfFrom Rule Scores Alone to Visual Plausibility Assessment (VLM-as-Judge):

10. A secondary, independent evaluation loop can be established where a powerful Vision-Language Model (VLM) acts as a judge. This VLM compares two generated layouts based on high-level functional quality criteria (e.g., Is the arrangement plausible for this room type?), providing an orthogonal signal to the rule scorer to catch subtle, non-geometric visual failures that hard rules miss, especially in complex scenarios like kitchens.

11. textbfFrom Simple Perturbations to Targeted Learning (Synthetic Pair DPO):

12. To ensure the model learns the precise effect of a single placement decision, the DPO training uses synthetic pairs where only one bounding box is perturbed while keeping the procedural trace identical. This localizes the preference signal to a single placement change, preventing reward hacking where models exploit superficial differences in trace style or surface form.

In summary, these improvements result in an AI system that generates layouts that are:

1. Functionally valid (passes hard geometric checks).

2. Editor-friendly (output is structured text/DSL).

3. Learned robustly (trained to follow explicit rules via DPO).

Furnished floor plans support real-estate visualization, interior design, and architectural workflows, yet automatic furnishing remains challenged by limited real-world data and the need to satisfy interacting geometric and functional constraints. We ask whether professional furnishing knowledge can be learned from real floor plans using a pretrained model, enabling direct constraint-aware layout generation without relying on costly iterative agentic inference. We introduce AntPlan, a curated dataset of 505 real professional architectural floor plans with dense furniture annotations spanning 92 object classes and ten residential room categories, and Architect-Ant, a framework for generating furniture layouts. Architect-Ant represents layouts with an editable coordinate-based DSL and first learns professional furnishing patterns through supervised fine-tuning. It is then optimized with GRPO using a Layout Rule Score (LRS) that aggregates geometric and functional constraints derived from professional plans, providing outcome-level supervision without prescribed reasoning traces. Experiments against diverse state-of-the-art baselines show that Architect-Ant combines low geometric violation rates with high functional completeness, while qualitative results more closely reflect real-world residential furnishing patterns. The resulting layouts remain object-level editable and can be converted into 3D scenes.